No more grid search! How to build models effectively by Thomas Huijskens

Confidential and proprietary: Any use of this material without
specific permission of McKinsey & Company is strictly prohibited
Intelligence.
Beautifully engineered.
No more grid search! How to build models effectively
Thomas Huijskens
London Driverless AI meet-up
1st August, 2019

2All content copyright © 2019 QuantumBlack, a McKinsey company
• Typically, we want to optimize some performance metric 𝑓ℳ for a model ℳ, on a hold-out set of data.
• The model ℳ has some hyperparameters 𝜃 that we need to specify, and the performance of ℳ is highly
dependent on the settings of 𝜃.
• Example: For a classification problem, ℳ could be a support vector machine, and the performance
metric 𝑓ℳ could be the out-of-sample AUC.
• Because the performance of the SVM depends on hyperparameters 𝜃, the metric 𝑓ℳ depends on 𝜃 as
well.
Most models have a number of hyperparameters that need to be chosen
a priori..

.. and our goal is to find the values of 𝜃 for which the value of the (out-
of-sample) performance 𝑓ℳ is optimal
𝑓ℳ 𝜃 𝐷)
Learner ℳ
Performance
function 𝑓
𝑓 depends on the hyperparameters 𝜃 of
the model ℳ
𝑓 is conditional on the
data 𝐷

Manual grid search works for low-dimensional problems, but does not
scale well to higher dimensions
Curse of dimensionality
makes this inefficient in
higher dimensional
problems

Random grid search works better in higher dimensions, but..

.. shouldn't previously evaluated hyperparameter values guide us in the
search for the true optimal value?
Low values of L here,
should probably not
focus in this area
High values of L here,
should probably focus
in this area

Advanced hyper-parameter
optimization methods

Advanced HPO methods are only beneficial in certain situations
In advanced optimisation methods, we spend more computational time to figure out what
set of parameters 𝜃 we are evaluating the performance metric 𝑓ℳ 𝜃 for.
As a result, applying these methods only really makes sense if
• The number of hyperparameters is very high (𝜃 is of high dimension); or
• It is computationally very expensive to evaluate 𝑓ℳ(𝜃) for a single point 𝜃.

There are now a wide range of hyper-parameter optimisation methods
HPO
Variant
Multi-
objective
Multi-fidelity
Search
Random / grid
search
Population-
based
Bayesian
optimisation
SMAC3: https://github.com/automl/SMAC3
Hyperopt: http://hyperopt.github.io/hyperopt/
Ax: https://github.com/facebook/Ax
Tune: https://ray.readthedocs.io/en/latest/tune.html
TPOT: https://github.com/EpistasisLab/tpot
Hyperband: https://github.com/thuijskens/scikit-hyperband
HpBandSter: https://github.com/automl/HpBandSter
Fabolas: https://github.com/automl/RoBO
MTBO: https://github.com/automl/RoBO

BO-HB is a recent method that combines Bayesian optimisation and
multi-fidelity methods
HPO
Variant
Multi-
objective
Multi-fidelity
Search
Random / grid
search
Population-
based
Bayesian
optimisation
SMAC3: https://github.com/automl/SMAC3
Hyperopt: http://hyperopt.github.io/hyperopt/
TPOT: https://github.com/EpistasisLab/tpot
Hyperband: https://github.com/thuijskens/scikit-hyperband
HpBandSter: https://github.com/automl/HpBandSter
Fabolas: https://github.com/automl/RoBO
MTBO: https://github.com/automl/RoBO

Hyperband

Hyperband iteratively runs the successive halving subroutine on model
configurations
BOHB: Robust and efficient hyperparameter optimization at scale: https://www.automl.org/blog_bohb/

Successive halving suffers from the exploitation versus exploration
trade-off
Given a total computational budget, the user has to decide whether
• To try many configurations with a small budget each; or
• To try only a few configurations with a larger budget.
Hyperband hedges for this trade-off by:
• Dividing the total computational budget into several combinations of number of configurations vs.
budget.
• Call successive halving as a subroutine for each combination.

Computational budget
Number of
configurations
Successive halving
iteration
Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization, Lisha Li, Kevin Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, Ameet Talwalkar, 2018. https://arxiv.org/abs/1603.06560

Hyperband’s hedging strategy is very competitive against random
search
BOHB: Robust and Efficient Hyperparameter Optimization at Scale, Stefan Falkner, Aaron Klein, Frank Hutter, 2018. https://arxiv.org/abs/1807.01774

Bayesian optimisation

The classic Bayesian optimization algorithm consists of three steps
Evaluate performance
of learner f with
parameters 𝜃
Update current belief
of loss surface of the
learner f
Choose 𝜃 that
maximises some
utility over the
current belief
𝑓 | 𝑦)*+
𝑦)*+ = 𝑓 𝜃)*+
𝜃)*+

The performance function 𝑓ℳ is modelled to form a current belief of
what the loss surface looks like..
of learner f with
parameters 𝜃
learner f
Choose 𝜃 that
maximises some
utility over the
current belief
𝑓 | 𝑦)*+
𝑦)*+ = 𝑓 𝜃)*+
𝜃)*+

..acquisition functions are used to formalize what constitutes a ”best
guess”..
of learner f with
parameters 𝜃
learner f
Choose 𝜃 that
maximises some
utility over the
current belief
𝑓 | 𝑦)*+
𝑦)*+ = 𝑓 𝜃)*+
𝜃)*+

The expected improvement acquisition trades off exploitation of known
optimal areas, versus exploration of unexplored areas of the loss
surface
Exploitation of
known optimal
space
Exploration of
unknown space
A Tutorial on Bayesian Optimization of Expensive Cost Functions, with Application to Active User Modeling and Hierarchical Reinforcement Learning, Eric Brochu, Vlad M. Cora, Nando de Freitas, 2010. https://arxiv.org/abs/1012.2599

.. and finally the performance is computed for suggested parameters
of learner f with
parameters 𝜃
learner f
Choose 𝜃 that
maximises some
utility over the
current belief
𝑓 | 𝑦)*+
𝑦)*+ = 𝑓 𝜃)*+
𝜃)*+

Bayesian optimization is competitive against random search for longer
training budgets

Bayesian optimisation works well for larger budgets, whereas
hyperband works better for small to medium budgets

BO-HB replaces random search in Hyperband with Bayesian
optimisation

What HPO method should I use in practice?
• If it is possible to define substantially cheaper version of the performance function 𝑓ℳ:
Bayesian optimisation-hyperband (BO-HB)
• If this is not possible and
• if all hyperparameters are real-valued and one can only afford a few dozen function
evaluations: Gaussian-process based Bayesian optimisation (SMAC, Spearmint)
• if the configuration space is large and/or conditional: Random forest-based Bayesian
optimisation (hyperopt)

No more grid search! How to build models effectively by Thomas Huijskens

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à No more grid search! How to build models effectively by Thomas Huijskens

Similaire à No more grid search! How to build models effectively by Thomas Huijskens (20)

Plus de Sri Ambati

Plus de Sri Ambati (20)

Dernier

Dernier (20)

No more grid search! How to build models effectively by Thomas Huijskens